Spoken Interface for Correcting Phoneme Recognition Errors in Learning of Unknown Words

抄録

This paper presents a novel method for learning of phoneme sequence for out-of-vocabulary (OOV) words. In the method, a user can correct mis-recognized phoneme sequence of an OOV word by making corrective utterances repeatedly. The originalities of this method are: 1) the correction is run in an interactive way, rather than in a batch way, which makes the correction more efficient and, 2) the correction is based on the open-begin-end dynamic programming matching (OBE-DPM) and generalized posterior probability (GPP), which enables a user to use a word segment in a corrective utterance. Comparative experimental results with a maximum likelihood based baseline method which is run in a batch processing showed that the proposed method achieved 96.8% and 79.1% in phoneme and word accuracies for learning new words, with less than seven corrective utterances, while the baseline method achieved only 87.7% and 31.8%. We also found that by using the proposed method, the correct phoneme sequences can be obtained within two corrective utterances for the most words in the experiments.